research

Pass

Audited by Gen Agent Trust Hub on Nov 13, 2026

Risk Level: SAFEPROMPT_INJECTIONCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [PROMPT_INJECTION]: The skill contains instructions to conceal technical data from the user during the handoff process.
  • Evidence: In references/output-template.md, the instructions state: "YAML readiness block goes in run.md — not customer-facing unless asked." This practice limits the user's ability to review the exact parameters and configurations being passed to downstream skills (like configure) unless they are technically savvy enough to request the raw file.
  • [COMMAND_EXECUTION]: The skill instructs the agent to execute Python modules that are located in an external repository (fw-ai-external/cookbook) rather than within the skill package itself.
  • Evidence: SKILL.md and references/starter-jobs.md reference commands such as python -m examples.sft.muse_glimmer_starter prepare and python -m examples.sft.example_datasets. While these are vendor-provided resources from the same author (fw-ai-external), they represent execution of code not explicitly bundled with the skill's static files.
  • [INDIRECT_PROMPT_INJECTION]: The skill acts as a data ingestion point that translates natural language user goals into structured YAML manifests for subsequent execution.
  • Ingestion points: User-provided task descriptions and problem statements (identified in SKILL.md and references/interview-questions.md).
  • Boundary markers: No explicit sanitization or escaping of user-provided strings is mentioned before they are interpolated into the YAML readiness package in run.md.
  • Capability inventory: The skill has the capability to write to the filesystem (fireworks-training-runs/<run-id>/run.md) and execute Python CLI tools based on the manifest content.
  • Sanitization: references/interview-questions.md mentions a redacted task_summary, but other fields like dataset_plan and eval_plan are generated based on user input, which could be manipulated to influence the downstream configure skill.
Audit Metadata
Risk Level
SAFE
Analyzed
Nov 13, 2026, 05:41 PM